# Solutions to Axler, Linear Algebra Done Right 2nd Ed.

Edvard Fagerholm
edvard.fagerholm@¦helsinki.ﬁ[gmail.com¦
Beware of errors. I read the book and solved the exercises during spring break (one
week), so the problems were solved in a hurry. However, if you do ﬁnd better or interesting
solutions to the problems, I’d still like to hear about them. Also please don’t put this on
the Internet to encourage copying homework solutions...
1 Vector Spaces
1. Assuming that C is a ﬁeld, write z = a + bi. Then we have 1/z = z/zz = z/[z[
2
.
Plugging in the numbers we get 1/(a + bi) = a/(a
2
+ b
2
) − bi/(a
2
+ b
2
) = c + di. A
straightforward calculation of (c+di)(a+bi) = 1 shows that this is indeed an inverse.
2. Just calculate ((1 +

3)/2)
3
.
3. We have v + (−v) = 0, so by the uniqueness of the additive inverse (prop. 1.3)
−(−v) = v.
4. Choose a ,= 0 and v ,= 0. Then assuming av = 0 we get v = a
−1
av = a
−1
0 = 0.
5. Denote the set in question by A in each part.
(a) Let v, w ∈ A, v = (x
1
, x
2
, x
3
), w = (y
1
, y
2
, y
3
). Then x
1
+ 2x
2
+ 3x
3
= 0 and
y
1
+ 2y
2
+ 3y
3
= 0, so that 0 = x
1
+ 2x
2
+ 3x
3
+ y
1
+ 2y
2
+ 3y
3
= (x
1
+ y
1
) +
2(x
2
+y
2
) +3(x
3
+y
3
), so v +w ∈ A. Similarly 0 = a0 = ax
1
+2ax
2
+3ay
3
, so
av ∈ A. Thus A is a subspace.
(b) This is not a subspace as 0 ,∈ A.
(c) We have that (1, 1, 0) ∈ A and (0, 0, 1) ∈ A, but (1, 1, 0)+(0, 0, 1) = (1, 1, 1) ,∈ A,
so A is not a subspace.
(d) Let (x
1
, x
2
, x
3
), (y
1
, y
2
, y
3
) ∈ A. If x
1
= 5x
3
and y
1
= 5y
3
, then ax
1
= 5ax
3
,
so a(x
1
, x
2
, x
3
) ∈ A. Similarly x
1
+ y
1
= 5(x
3
+ y
3
), so that (x
1
, x
2
, x
3
) +
(y
1
, y
2
, y
3
) ∈ A. Thus A is a subspace.
1
6. Set U = Z
2
.
7. The set ¦(x, x) ∈ R
2
[ x ∈ R¦ ∪¦(x, −x) ∈ R
2
[ x ∈ R¦ is closed under multiplication
but is trivially not a subspace ((x, x) + (x, −x) = (2x, 0) doesn’t belong to it unless
x = 0).
8. Let ¦V
i
¦ be a collection of subspaces of V . Set U = ∩
i
V
i
. Then if u, v ∈ U. We
have that u, v ∈ V
i
for all i because V
i
is a subspace. Thus au ∈ V
i
for all i, so that
av ∈ U. Similarly u + v ∈ V
i
for all i, so u + v ∈ U.
9. Let U, W ⊂ V be subspaces. Clearly if U ⊂ W or W ⊂ U, then U ∪ W is clearly a
subspace. Assume then that U ,⊂ W and W ,⊂ U. Then we can choose u ∈ U ¸ W
and w ∈ W ¸ U. Assuming that U ∪ W is a subspace we have u + w ∈ U ∪ W.
Assuming that u + w ⊂ U we get w = u + w −u ⊂ U. Contradiction. Similarly for
u + w ∈ W. Thus U ∪ W is not a subspace.
10. Clearly U = U + U as U is closed under addition.
11. Yes and yes. Follows directly from commutativity and associativity of vector addition.
12. The zero subspace, ¦0¦, is clearly an additive identity. Assuming that we have
inverses, then the whole space V should have an inverse U such that U + V = ¦0¦.
Since V + U = V this is clearly impossible unless V is the trivial vector space.
13. Let W = R
2
. Then for any two subspaces U
1
, U
2
of W we have U
1
+ W = U
2
+ W,
so the statement is clearly false in general.
14. Let W = ¦p ∈ T(F) [ p =

n
i=0
a
i
x
i
, a
2
= a
5
= 0¦.
15. Let V = R
2
. Let W = ¦(x, 0) ∈ R
2
[ x ∈ R¦. Set U
1
= ¦(x, x) ∈ R
2
[ x ∈ R¦,
U
2
= ¦(x, −x) ∈ R
2
[ x ∈ R¦. Then it’s easy to see that U
1
+ W = U
2
+ W = R
2
,
but U
1
,= U
2
, so the statement is false.
2 Finite-Dimensional Vector Spaces
1. Let u
n
= v
n
and u
i
= v
i
+v
i+1
, i = 1, . . . , n−1. Now we see that v
i
=

n
j=i
u
i
. Thus
v
i
∈ span(u
1
, . . . , u
n
), so V = span(v
1
, . . . , v
n
) ⊂ span(u
1
, . . . , u
n
).
2. From the previous exercise we know that the span is V . As (v
1
, . . . , v
n
) is a linearly
independent spanning list of vectors we know that dimV = n. The claim now follows
from proposition 2.16.
2
3. If (v
1
+ w, . . . , v
n
+ w) is linearly dependent, then we can write
0 = a
1
(v
1
+ w) + . . . + a
n
(v
n
+ w) =
n

i=1
a
i
v
i
+ w
n

i=1
a
i
,
where a
i
,= 0 for some i. Now

n
i=1
a
i
,= 0 because otherwise we would get
0 =
n

i=1
a
i
v
i
+ w
n

i=1
a
i
=
n

i=1
a
i
v
i
and by the linear independence of (v
i
) we would get a
i
assumption. Thus
w = −
_
n

i=1
a
i
_
−1
n

i=1
a
i
v
i
,
so w ∈ span(v
1
, . . . , v
n
).
4. Yes, multiplying with a nonzero constant doesn’t change the degree of a polynomial
and adding two polynomials either keeps the degree constant or makes it the zero
polynomial.
5. (1, 0, . . .), (0, 1, 0, . . .), (0, 0, 1, 0, . . .), . . . is trivially linearly independent. Thus F

isn’t ﬁnite-dimensional.
6. Clearly T(F) is a subspace which is inﬁnite-dimensional.
7. Choose v
1
∈ V . Assuming that V is inﬁnite-dimensional span(v
1
) ,= V . Thus we can
choose v
2
∈ V ¸ span(v
1
). Now continue inductively.
8. (3, 1, 0, 0, 0), (0, 0, 7, 1, 0), (0, 0, 0, 0, 1) is clearly linearly independent. Let (x
1
, x
2
, x
3
, x
4
, x
5
) ∈
U. Then (x
1
, x
2
, x
3
, x
4
, x
5
) = (3x
2
, x
2
, 7x
4
, x
4
, x
5
) = x
2
(3, 1, 0, 0, 0)+x
4
(0, 0, 7, 1, 0)+
x
5
(0, 0, 0, 0, 1), so the vectors span U. Hence, they form a basis.
9. Choose the polynomials (1, x, x
2
−x
3
, x
3
). They clearly span T
3
(F) and form a basis
by proposition 2.16.
10. Choose a basis (v
1
, . . . , v
n
). Let U
i
= span(v
i
). Now clearly V = U
1
⊕ . . . ⊕ U
n
by
proposition 2.19.
11. Let dimU = n = dimV . Choose a basis (u
1
, . . . , u
n
) for U and extend it to a
basis (u
1
, . . . , u
n
, v
1
, . . . , v
k
) for V . By assumption a basis for V has length n, so
(u
1
, . . . , u
n
) spans V .
3
12. Let U = span(u
0
, . . . , u
m
). As U ,= T
m
(F) we have by the previous exercise that
dimU < dimT
m
(F) = m + 1. As (p
0
, . . . , p
m
) is a spanning list of vectors having
length m + 1 it is not linearly independent.
13. By theorem 2.18 we have 8 = dimR
8
= dimU + dimW − dim(U ∩ W) = 4 + 4 −
dim(U ∩W). Since U +W = R
8
, we have that dim(U ∩W) = 0 and the claim follows
14. Assuming that U ∩ W = ¦0¦ we have by theorem 2.18 that
9 = dimR
9
= dimU + dimW −dim(U ∩ W) = 5 + 5 −0 = 10.
15. Let U
1
= ¦(x, 0) ∈ R
2
[ x ∈ R¦, U
2
= ¦(0, x) ∈ R
2
[ x ∈ R¦, U
3
= ¦(x, x) ∈ R
2
[ x ∈
R¦. Then U
1
∩ U
2
= ¦0¦, U
1
∩ U
3
= ¦0¦ and U
2
∩ U
3
= ¦0¦. Thus the left-hand side
of the equation in the problem is 2 while the right-hand side is 3. Thus we have a
counterexample.
16. Choose a basis (u
i
1
, . . . , u
i
n
i
) for each U
i
. Then the list of vectors (u
1
1
, . . . , u
m
nm
) has
length dimU
1
+ . . . + dimU
m
and clearly spans U
1
+ . . . + U
m
proving the claim.
17. By assumption the list of vectors (u
1
1
, . . . , u
m
nm
) from the proof of the previous exercise
is linearly independent. Since V = U
1
+. . .+U
n
= span(u
1
1
, . . . , u
m
nm
) the claim follows.
3 Linear Maps
1. Any v ,= 0 spans V . Thus, if w ∈ V , we have w = av for some a ∈ F. Now T(v) = av
for some a ∈ F, so for w = bv ∈ V we have T(w) = T(bv) = bT(v) = bav = aw. Thus
T is multiplication by a scalar.
2. Deﬁne f by e.g.
f(x, y) =
_
x, x = y
0, x ,= y
,
then clearly f satisﬁes the condition, but is non-linear.
3. To deﬁne a linear map it’s enough to deﬁne the image of the elements of a basis.
Choose a basis (u
1
, . . . , u
m
) for U and extend it to a basis (u
1
, . . . , u
m
, v
1
, . . . , v
n
) of
V . If S ∈ L(U, W). Choose some vectors w
1
, . . . , w
n
∈ W. Now deﬁne T ∈ L(V, W)
by T(u
i
) = S(u
i
), i = 1, . . . , m and T(v
i
) = w
i
, i = 1, . . . , n. These relations deﬁne
the linear map and clearly T
|U
= S.
4. By theorem 3.4 dimV = dimnull T +dimrange T. If u ,∈ null T, then dimrange T >
0, but as dimrange T ≤ dimF = 1 we have dimrange T = 1 and it follows that
4
dimnull T = dimV −1. Now choose a basis (u
1
, . . . , u
n
) for null T. By our assumption
(u
1
, . . . , u
n
, u) is linearly independent and has length dimV . Hence, it’s a basis for
V . This implies that
V = span(u
1
, . . . , u
n
, u) = span(u
1
, . . . , u
n
) ⊕span(u) = null T ⊕¦au [ a ∈ F¦.
5. By linearity 0 =

n
i=1
a
i
T(v
i
) =

n
i=1
T(a
i
v
i
) = T(

n
i=1
a
i
v
i
). By injectivity we
have

n
i=1
a
i
v
i
= 0, so that a
i
= 0 for all i. Hence (T(v
1
), . . . , T(v
n
)) is linearly
independent.
6. If n = 1 the claim is trivially true. Assume that the claim is true for n = k.
If S
1
, . . . , S
k+1
satisfy the assumptions, then S
1
S
k
is injective by the induction
hypothesis. Let T = S
1
S
k
. If u ,= 0, then by injectivity of S
k+1
we S
k+1
u ,= 0
and by injectivity of T we have TS
k+1
u ,= 0. Hence TS
k+1
is injective and the claim
follows.
7. Let w ∈ W. By surjectivity of T we can ﬁnd a vector v ∈ V such that T(v) = w.
Writing v = a
1
v
1
+ . . . + a
n
v
n
we get w = T(v) = T(a
1
v
1
+ . . . + a
n
v
n
) = a
1
T(v
1
) +
. . . + a
n
T(v
n
) proving the claim.
8. Let (u
1
, . . . , u
n
) be a basis for null T and extend it to a basis (u
1
, . . . , u
n
, v
1
, . . . , v
k
)
of V . Let U := span(v
1
, . . . , v
k
). Then by construction null T ∩ U = ¦0¦ and for an
arbitrary v ∈ V we have that v = a
1
u
1
+ . . . + a
n
u
n
+ b
1
v
n
+ . . . + b
k
v
k
, so that
T(v) = b
1
T(v
1
) + . . . + b
k
T(v
k
).
Hence range T = T(U).
9. It’s easy to see that (5, 1, 0, 0), (0, 0, 7, 1) is a basis for null T (see exercise 2.8). Hence
dimrange T = dimF
4
− dimnull T = 4 − 2 = 2. Thus range T = F
2
, so that T is
surjective.
10. It’s again easy to see that (3, 1, 0, 0, 0), (0, 0, 1, 1, 1) is a basis of null T. Hence, we get
dimrange T = dimF
5
−dimnull T = 5 −2 = 3 which is impossible.
11. This follows trivially from dimV = dimnull T + dimrange T.
12. From dimV = dimnull T +dimrange T it follows trivially that if we have a surjective
linear map T ∈ L(V, W), then dimV ≥ dimW. Assume then that dimV ≥ dimW.
Choose a basis (v
1
, . . . , v
n
) for V and a basis (w
1
, . . . , w
m
) for W. We can deﬁne
a linear map T ∈ L(V, W) by letting T(v
i
) = w
i
, i = 1, . . . , m, T(v
i
) = 0, i =
m + 1, . . . , n. Clearly T is surjective.
5
13. We have that dimrange T ≤ dimW. Thus we get dimnull T = dimV −dimrange T ≥
dimV −dimW. Choose an arbitrary subspace U of V of such that dimU ≥ dimV −
dimW. Let (u
1
, . . . , u
n
) be a basis of U and extend it to a basis (u
1
, . . . , u
n
, v
1
, . . . , v
k
)
of V . Now we know that k ≤ dimW, so let (w
1
, . . . , w
m
) be a basis for W. Deﬁne
a linear map T ∈ L(V, W) by T(u
i
) = 0, i = 1, . . . , n and T(v
i
) = w
i
, i = 1, . . . , k.
Clearly null T = U.
14. Clearly if we can ﬁnd such a S, then T is injective. Assume then that T is injective.
Let (v
1
, . . . , v
n
) be a basis of V . By exercise 5 (T(v
1
), . . . , T(v
n
)) is linearly inde-
pendent so we can extend it to a basis (T(v
1
), . . . , T(v
n
), w
1
, . . . , w
m
) of W. Deﬁne
S ∈ L(W, V ) by S(T(v
i
)) = v
i
, i = 1, . . . , n and S(w
i
) = 0, i = 1, . . . , m. Clearly
ST = I
V
.
15. Clearly if we can ﬁnd such a S, then T is surjective. Otherwise let (v
1
, . . . , v
n
) be a
basis of V . By assumption (T(w
1
), . . . , T(w
n
)) spans W. By the linear dependence
lemma we can make the list of vectors (T(w
1
), . . . , T(w
n
)) a basis by removing some
vectors. Without loss of generality we can assume that the ﬁrst m vectors form the
basis (just permute the indices). Thus T(w
1
), . . . , T(w
m
)) is a basis of W. Deﬁne
the map S ∈ L(W, V ) by S(T(w
i
)) = w
i
, i = 1, . . . , m. Now clearly TS = I
W
.
16. We have that dimU = dimnull T + dimrange T ≤ dimnull T + dimV . Substituting
dimV = dimnull S + dimrange S and dimU = dimnull ST + dimrange ST we get
dimnull ST + dimrange ST ≤ dimnull T + dimnull S + dimrange S.
Clearly dimrange ST ≤ dimrange S, so our claim follows.
17. This is nothing but pencil pushing. Just take arbitrary matrices satisfying the re-
quired dimensions and calculate each expression and the equalities easily fall out.
18. Ditto.
19. Let V = (x
1
, . . . ,
n
). From proposition 3.14 we have that
/(Tv) = /(T)/(v) =
_
¸
_
a
1,1
a
1,n
.
.
.
.
.
.
.
.
.
a
m,1
a
m,n
_
¸
_
_
¸
_
x
1
.
.
.
x
n
_
¸
_ =
_
¸
_
a
1,1
x
1
+ . . . + a
1,n
x
n
.
.
.
a
m,1
x
1
+ . . . + a
m,n
x
n
_
¸
_
which shows that Tv = (a
1,1
x
1
+ . . . + a
1,n
x
n
, . . . , a
m,1
x
1
+ . . . + a
m,n
x
n
).
20. Clearly dimMat(n, 1, F) = n = dimV . We have that Tv = 0 if and only if v =
0v
1
+ . . . , 0v
n
= 0, so null T = ¦0¦. Thus T is injective and hence invertible.
6
21. Let e
i
denote the n 1 matrix with a 1 in the ith row and 0 everywhere else. Let T
be the linear map. Deﬁne a matrix A := [T(e
1
), . . . , T(e
n
)]. Now it’s trivial to verify
that
Ae
i
= T(e
i
).
By distributivity of matrix multiplication (exercise 17) we get for an arbitrary v =
a
1
e
1
+ . . . + a
n
e
n
that
Av = A(a
1
e
1
+ . . . + a
n
e
n
) = a
1
Ae
1
+ . . . + a
n
Ae
n
= a
1
T(e
1
) + . . . + a
n
T(e
n
) = T(a
1
e
1
+ . . . + a
n
e
n
),
so the claim follows.
22. From theorem 3.21 we have ST invertible ⇔ ST bijective ⇔ S and T bijective ⇔ S
and T invertible.
23. By symmetry it’s suﬃcient to prove this in one direction only. Thus if TS = I. Then
ST(Su) = S(TSu) = Su for all u. As S is bijective Su goes through the whole space
V as u varies, so ST = I.
24. Clearly TS = ST if T is a scalar multiple of the identity. For the other direction
assume that TS = ST for every linear map S ∈ L(V ).
25. Let T ∈ L(F
2
) be the operator T(a, b) = (a, 0) and S ∈ L(F
2
) the operator S(a, b) =
(0, b). Then neither one is injective and hence invertible by 3.21. However, T + S is
the identity operator which is trivially invertible. This can be trivially generalized
to spaces arbitrary spaces of dimension ≥ 2.
26. Write
A :=
_
¸
_
a
1,1
a
1,n
.
.
.
.
.
.
.
.
.
a
n,1
a
n,n
_
¸
_, x =
_
¸
_
x
1
.
.
.
x
n
_
¸
_.
Then the system of equations in a) reduces to Ax = 0. Now A deﬁnes a linear map
from Mat(n, 1, F) to Mat(n, 1, F). What a) states is now that the map is injective
while b) states that it is surjective. By theorem 3.21 these are equivalent.
4 Polynomials
1. Let λ
1
, . . . , λ
m
be m distinct numbers and k
1
, . . . , k
m
> 0 such that their sum is n.
Then

m
i=1
(x −λ
i
)
k
i
is clearly such a polynomial.
7
2. Let p
i
(x) =

j=i
(x − z
j
), so that deg p
i
= m. Now we have that p
i
(z
j
) ,= 0 if and
only if i = j. Let c
i
= p
i
(z
i
). Deﬁne
p(x) =
m+1

i=1
w
i
c
−1
i
p
i
(x).
Now deg p = m and p clearly p(z
i
) = w
i
.
3. We only need to prove uniqueness as existence is theorem 4.5. Let s

, r

be other such
polynomials. Then we get that
0 = (s −s

)p + (r −r

) ⇔ r

−r = (s −s

)p
We know that for any polynomials p ,= 0 ,= q we have deg pq = deg p + deg q.
Assuming that s ,= s

we have deg(r

− r) < deg(s − s

)p which is impossible. Thus
s = s

implying r = r

.
4. Let λ be a root of p. Thus we can write p(x) = (x −λ)q(x). Now we get
p

(x) = q(x) + (x −λ)q

(x).
Thus λ is a root of p

if and only if λ is a root of q i.e. λ is a multiple root. The
statement follows.
5. Let p(x) =

n
i=0
a
i
x
i
, where a
i
∈ R. By the fundamental theorem of calculus we
have a complex root z. By proposition 4.10 z is also a root of p. Then z and z are
roots of equal multiplicity (divide by (x − z)(x − z) which is real). Thus if we have
no real roots we have an even number of roots counting multiplicity, but the number
of roots counting multiplicity is deg p hence odd. Thus p has a real root.
5 Eigenvalues and Eigenvectors
1. An arbitrary element of U
1
+. . . +U
n
is of the form u
1
+. . . +u
n
where u
i
∈ U
i
. Thus
we get T(u
1
+ . . . + u
n
) = T(u
1
) + . . . + T(u
n
) ∈ U
1
+ . . . + U
n
by the assumption
that T(u
i
) ∈ U
i
for all i.
2. Let V = ∩
i
U
i
where U
i
is invariant under T for all i. Let v ∈ V , so that v ∈ U
i
for all
i. Now T(v) ∈ U
i
for all i by assumption, so T(v) ∈ ∩U
i
= V . Thus V is invariant
under T.
3. The clame is clearly true for U = ¦0¦ or U = V . Assume that ¦0¦ ,= U ,= V . Let
(u
1
, . . . , u
n
) be a basis for U and extend it to a basis (u
1
, . . . , u
n
, v
1
, . . . , v
m
). By
our assumption m ≥ 1. Deﬁne a linear operator by T(u
i
) = v
1
, i = 1, . . . , n and
T(v
i
) = v
1
, i = 1, . . . , m. Then clearly U is not invariant under T.
8
4. Let u ∈ null(T −λI), so that T(u) −λu = 0. ST = TS gives us
0 = S(Tu −λu) = STu −λSu = TSu −λSu = (T −λI)Su,
so that Su ∈ null(T −λI).
5. Clearly T(1, 1) = (1, 1), so that 1 is an eigenvalue. Also T(−1, 1) = (1, −1) =
−(−1, 1) so −1 is another eigenvalue. By corollary 5.9 these are all eigenvalues.
6. We easily see that T(0,0,1)=(0,0,5), so that 5 is an eigenvalue. Also T(1, 0, 0) =
(0, 0, 0) so 0 is an eigenvalue. Assume that λ ,= 0 and T(z
1
, z
2
, z
3
) = (2z
2
, 0, 5z
3
) =
λ(z
1
, z
2
, z
3
). From the assumption λ ,= 0 we get z
2
= 0, so the equation is of the
form (0, 0, 5z
3
) = λ(z
1
, 0, z
3
). Again we see that z
1
= 0 so we get the equation
(0, 0, 5z
3
) = λ(0, 0, z
3
). Thus 5 is the only non-zero eigenvalue.
7. Notice that the range of T is the subspace ¦(x, . . . , x) ∈ F
n
[ x ∈ F¦ and has dimension
1. Thus dimrange T = 1, so dimnull T = n − 1. Assume that T has two distinct
eigenvalues λ
1
, λ
2
and assume that λ
1
,= 0 ,= λ
2
. Let v
1
, v
2
be the corresponding
eigenvectors, so by theorem 5.6 they are linearly independent. Then v
1
, v
2
,∈ null T,
but dimnull T = n − 1, so this is impossible. Hence T has at most one non-zero
eigenvalue hence at most two eigenvalues.
Because T is not injective we know that 0 is an eigenvalue. We also see that
T(1, . . . , 1) = n(1, . . . , 1), so n is another eigenvalue. By the previous paragraph,
these are all eigenvalues of T.
8. Let a ∈ F. Now we have T(a, a
2
, a
3
, . . .) = (a
2
, a
3
, . . .) = a(a, a
2
, . . .), so every a ∈ F
is an eigenvalue.
9. Assume that T has k +2 distinct eigenvalues λ
1
, . . . , λ
k+2
with corresponding eiven-
vectors v
1
, . . . , v
k+2
. By theorem 5.6 these eigenvectors are linearly independent. Now
Tv
i
= λ
i
v
i
and dimspan(Tv
1
, . . . , Tv
k+2
) = dimspan(λ
1
v
1
, . . . , λ
k+2
v
k+2
) ≥ k + 1
(it’s k + 2 if all λ
i
are non-zero, otherwise k + 1). This is a contradiction as
dimrange T = k and span(λ
1
v
1
, . . . , λ
k+2
v
k+2
) ⊂ range T.
10. As T = (T
−1
)
−1
we only need to show this in one direction. If T is invertible, then
0 is not an eigenvalue. Now let λ be an eigenvalue of T and v the corresponding
eigenvector. From Tv = λv we get that T
−1
λv = λT
−1
v = v, so that T
−1
v = λ
−1
v.
11. Let λ be an eigenvalue of TS and v the corresponding eigenvector. Then we get
STSv = Sλv = λSv, so if Sv ,= 0, then it is is an eigenvector for the eigenvalue λ. If
Sv = 0, then TSv = 0, so λ = 0. As Sv = 0 we know that S is not injective, so ST
is not injective and it has eigenvalue 0. Thus if λ is an eigenvalue of TS, then it’s an
eigenvalue of ST. The other implication follows by symmetry.
9
12. Let (v
1
, . . . , v
n
) be a basis of V . By assumption (Tv
1
, . . . , Tv
n
) = (λ
1
v, . . . , λ
n
v
n
).
We need to show that λ
i
= λ
j
for all i, j. Choose i ,= j, then by our assumption
v
i
+ v
j
is an eigenvector, so T(v
i
+ v
j
) = λ
i
v
i
+ λ
j
v
j
= λ(v
i
+ v
j
) = λv
i
+ λv
j
. This
means that (λ
i
− λ)v
i
+ (λ
j
− λ)v
j
= 0. Because (v
i
, v
j
) is linearly independent we
get that λ
i
−λ = 0 = λ
j
−λ i.e. λ
i
= λ
j
.
13. Let v
1
∈ V and extend it to a basis (v
1
, . . . , v
n
) of V . Now Tv
1
= a
1
v
1
+ . . . + a
n
v
n
.
Let U
i
be the subspace generated by the vectors v
j
, j ,= i. By our assumption each
U
i
is an invariant subspace. Let Tv
1
= a
1
v
1
+ . . . + a
n
v
n
. Now v
1
∈ U
i
for i > 1. So
let j > 1, then Tv ∈ U
j
imples a
j
= 0. Thus Tv
1
= a
1
v
1
. We see that v
1
was an
eigenvector. The result now follows from the previous exercise.
14. Clearly (STS
−1
)
n
= ST
n
S
−1
, so
n

i=0
a
i
(STS
−1
)
i
=
n

i=0
a
i
ST
i
S
−1
= S
_
n

i=0
a
i
T
i
_
S
−1
.
15. Let λ be an eigenvalue of T and v ∈ V a corresponding eigenvector. Let p(x) =

n
i=0
a
i
x
i
. Then we have
p(T)v =
n

i=0
a
i
T
i
v =
n

i=0
a
i
λ
i
v = p(λ)v.
Thus p(λ) is an eigenvalue of p(T). Then let a be an eigenvalue of p(T) and v ∈ V
the corresponding eigenvector. Let q(x) = p(x) −a. By the fundamental theorem of
algebra we can write q(x) = c

n
i=1
(x −λ
i
). Now q(T)v = 0 and q(T) = c

n
i=1
(T −
λ
i
I). As q(T) is non-injective we have that T −λ
i
I is non-injective for some i. Hence
λ
i
is an eigenvalue of T. Thus we get 0 = q(λ
i
) = p(λ
i
) −a, so that a = p(λ
i
).
16. Let T ∈ L(R
2
) be the map T(x, y) = (−y, x). On page 78 it was shown that it has
no eigenvalue. However, T
2
(x, y) = T(−y, x) = (−x, −y), so −1 it has an eigenvalue.
17. By theorem 5.13 T has an upper-triangular matrix with respect to some basis (v
1
, . . . , v
n
).
The claim now follows from proposition 5.12.
18. Let T ∈ L(F
2
) be the operator T(a, b) = (b, a) with respect to the standard basis.
Now T
2
= I, so T is clearly invertible. However, T has the matrix
_
0 1
1 0
_
.
10
19. Take the operator T in exercise 7. It is clearly not invertible, but has the matrix
_
¸
_
1 1
.
.
.
.
.
.
.
.
.
1 1
_
¸
_.
20. By theorem 5.6 we can choose basis (v
1
, . . . , v
n
) for V where v
i
is an eigenvector of
T corresponding to the eigenvalue λ
i
. Let λ be the eigenvalue of S corresponding to
v
i
. Then we get
STv
i
= Sλ
i
v
i
= λ
i
Sv
i
= λ
i
λv
i
= λλ
i
v
i
= λTv
i
= Tλv
i
= TSv
i
,
so ST and TS agree on a basis of V . Hence they are equal.
21. Clearly 0 is an eigenvalue and the corresponding eigenvectors are null T = W ¸ ¦0¦.
Now assume λ ,= 0 is an eigenvalue and v = u + w is a corresponding eigenvector.
Then P
U,W
(u + w) = u = λu + λw ⇔ (1 −λ)u = λw. Thus λw ∈ U, so λw = 0, but
λ ,= 0, so we have w = 0 which implies (1 −λ)u = 0. Because v is an eigenvector we
have v = u + w = u ,= 0, so that 1 − λ = 0. Hence λ = 1. As we can choose freely
our u ∈ U, so that it’s non-zero we see that the eigenvectors corresponding to 1 are
u = U ¸ ¦0¦.
22. As dimV = dimnull P + dimrange P, it’s clearly suﬃcient to prove that null P ∩
range P = ¦0¦. Let v ∈ null P ∩ range P. As v ∈ range P, we can ﬁnd a u ∈ V such
that Pu = v. Thus we have that v = Pu = P
2
u = Pv, but v ∈ null P, so Pv = 0.
Thus v = 0.
23. Let T(a, b, c, d) = (−b, a, −d, c), then T is clearly injective, so 0 is not an eigenvalue.
If λ ,= 0 is an eigenvalue and λ(a, b, c, d) = (−b, a, −d, c). We clearly see that a ,=
0 ⇔ b ,= 0 and similarly with c, d. By symmetry we can assume that a ,= 0 (an
eigenvector is non-zero). Then we have λa = −b and λb = a. Substituting we get
λ
2
b = −b ⇔ (λ
2
+ 1)b = 0. As b ,= 0 we have λ
2
+ 1 = 0, but this equation has no
solutions in R. Hence T has no eigenvalue.
24. If U is an invariant subspace of odd degree, then by theorem 5.26 T
|U
has an eigenvalue
λ with eigenvector v. Then λ is an eigenvalue of T with eigenvector v against our
assumption. Thus T has no subspace of odd degree.
6 Inner-Product Spaces
1. From the law of cosines we get |x − y|
2
= |x|
2
+ |y|
2
− 2|x||y| cos θ. Solving we
get
|x||y| cos θ =
|x|
2
+|y|
2
−|x −y|
2
2
.
11
Now let x = (x
1
, x2), y = (y
1
, y
2
). A straight calculation shows that
|x|
2
+|y|
2
−|x−y|
2
= x
2
1
+x
2
2
+y
2
1
+y
2
2
−(x
1
−y
1
)
2
−(x
2
−y
2
)
2
= 2(x
1
y
1
+x
2
y
2
),
so that
|x||y| cos θ =
|x|
2
+|y|
2
−|x −y|
2
2
=
2(x
1
y
1
+ x
2
y
2
)
2
= ¸x, y) .
2. If ¸u, v) = 0, then ¸u, av) = 0, so by the Pythagorean theorem |u| ≤ |u| + |av| =
|u+av|. Then assume that |u| ≤ |u+av| for all a ∈ F. We have |u|
2
≤ |u+av|
2

¸u, u) ≤ ¸u + av, u + av). Thus we get
¸u, u) ≤ ¸u, u) +¸u, av) +¸av, u) +¸av, av) ,
so that
−2Re a ¸u, v) ≤ [a[
2
|v|
2
.
Choose a = −t ¸u, v) with t > 0, so that
2t ¸u, v)
2
≤ t
2
¸u, v)
2
|v|
2
⇔ 2 ¸u, v)
2
≤ t ¸u, v)
2
|v|
2
.
If v = 0, then clearly ¸u, v) = 0. If not choose t = 1/|v|
2
, so that we get
2 ¸u, v)
2
≤ ¸u, v)
2
.
Thus ¸u, v) = 0.
3. Let a = (a
1
,

2a
2
, . . . ,

na
n
) ∈ R
n
and b = (b
1
, b
2
/

2, . . . , b
n
/

n) ∈ R
n
. This
equality is then simply ¸a, b)
2
≤ |a|
2
|b|
2
which follows directly from the Cauchy-
Schwarz inequality.
4. From the parallelogram equality we have |u + v|
2
+ |u − v|
2
= 2(|u|
2
+ |v|
2
).
Solving for |v| we get |v| =

17.
5. Set e.g. u = (1, 0), v = (0, 1). Then |u| = 1, |v| = 1, |u + v| = 2, |u − v| = 2.
Assuming that the norm is induced by an inner-product, we would have by the
parallelogram inequality
8 = 2
2
+ 2
2
= 2(1
2
+ 1
2
) = 4,
which is clearly false.
6. Just use |u| = ¸u, u) and simplify.
7. See previous exercise.
12
8. This exercise is a lot trickier than it might seem. I’ll prove it for R, the proof for C
is almost identical except for the calculations that are longer and more tedious. All
norms on a ﬁnite-dimensional real vector space are equivalent. This gives us
lim
n→∞
|r
n
x + y| = |λx + y|
when r
n
→ λ. This probably doesn’t make any sense, so check a book on topology
or functional analysis or just assume the result.
Deﬁne ¸u, v) by
¸u, v) =
|u + v|
2
−|u −v|
2
4
.
Trivially we have |u|
2
= ¸u, u), so positivity deﬁniteness follows. Now
4(¸u + v, w) −¸u, w) −¸v, w)) = |u + v + w|
2
−|u + v −w|
2
−(|u + w|
2
−|u −w|
2
)
−(|v + w|
2
−|v −w|
2
)
= |u + v + w|
2
−|u + v −w|
2
+(|u −w|
2
+|v −w|
2
)
−(|u + w|
2
+|v + w|
2
)
We can apply the parallelogram equality to the two parenthesis getting
4(¸u + v, w) −¸u, w) −¸v, w)) = |u + v + w|
2
−|u + v −w|
2
+
1
2
(|u + w −2w|
2
+[u −v[
2
)

1
2
(|u + v + 2w|
2
+|u −v|
2
)
= |u + v + w|
2
−|u + v −w|
2
+
1
2
|u + w −2w|
2

1
2
|u + v + 2w|
2
= (|u + v + w|
2
+|w|
2
) +
1
2
|u + v −2w|
2
− (|u + v −w|
2
+|w|
2
) −
1
2
|u + v + 2w|
2
Applying the parallelogram equality to the two parenthesis we and simplifying we
are left with 0. Hence we get additivity in the ﬁrst slot. It’s easy to see that
¸−u, v) = −¸u, v), so we get ¸nu, v) = n¸u, v) for all n ∈ Z. Now we get
¸u, v) = ¸nu/n, v) = n¸u/n, v) ,
13
so that ¸u/n, v) = 1/n¸u, v). Thus we have ¸nu/m, v) = n/m¸u, v). It follows that
we have homogeneity in the ﬁrst slot when the scalar is rational. Now let λ ∈ R and
choose a sequence (r
n
) of rational numbers such that r
n
→ λ. This gives us
λ¸u, v) = lim
n→∞
r
n
¸u, v) = lim
n→∞
¸r
n
u, v)
= lim
n→∞
1
4
(|r
n
u + v|
2
−|r
n
u −v|
2
)
=
1
4
(|λu + v|
2
−|λu −v|
2
)
= ¸λu, v)
Thus we have homogeneity in the ﬁrst slot. We trivially also have symmetry, so we
have an inner-product. The proof for F = C can be done by deﬁning the inner-product
from the complex polarizing identity. And by using the identities
¸u, v)
0
=
1
4
(|u + v|
2
−|u −v|
2
) ⇒ ¸u, v) = ¸u, v)
0
+¸u, iv)
0
i.
and using the properties just proved for ¸, )
0
.
9. This is really an exercise in calculus. We have integrals of the types
¸sinnx, sinmx) =
_
π
−π
sinnxsinmxdx
¸cos nx, cos mx) =
_
π
−π
cos nxcos mxdx
¸sinnx, cos mx) =
_
π
−π
sinnxcos mxdx
and they can be evaluated using the trigonometric identities
sinnxsinmx =
cos((n −m)x) −cos((n + m)x)
2
cos nxcos mx =
cos((n −m)x) + cos((n + m)x)
2
cos nxsinmx =
sin((n −m)x) −sin((n + m)x)
2
.
10. e
1
= 1, then e
2
=
x−x,11
x−x,11
. Where
¸x, 1) =
_
1
0
x = 1/2,
14
so that
|x −1/2| = ¸x −1/2, x −1/2) =
¸
_
1
0
(x −1/2)
2
dx =
_
1/12,
which gives e
2
=

12(x − 1/2) =

3(2x − 1). Then to continue pencil pushing we
have
¸
x
2
, 1
_
= 1/3,
_
x
2
,

3(2x −1)
_
=

3/6,
so that x
2

¸
x
2
, 1
_
1 −
¸
x
2
,

3(2x −1)
_ √
3(2x −1) = x
2
−1/3 −1/2(2x −1). Now
|x
2
−1/3 −1/2(2x −1)| = 1/(6

5)
giving e
3
=

5(6x
2
−6x+1). The orthonormal basis is thus (1,

3(2x−1),

5(6x
2

6x + 1).
11. Let (v
1
, . . . , v
n
) be a linearly dependent list. We can assume that (v
1
, . . . , v
n−1
) is
linearly independent and v
n
∈ span(v
1
, . . . , v
n
). Let P denote the projection on the
subspace spanned by (v
1
, . . . , v
n−1
). Then calculating e
n
with the Gram-Schmidt
algorithm gives us
e
n
=
v
n
−P(v
n
)
|v
n
−Pv
n
|
,
but v
n
= Pv
n
by our assumption, so e
n
= 0. Thus the resulting list will simply
contain zero vectors. It’s easy to see that we can extend the algorithm to work for
linearly dependent lists by tossing away resulting zero vectors.
12. Let (e
1
, . . . , e
n−1
) be an orthogonal list of vectors. Assume that (e
1
, . . . , e
n
) and
(e
1
, . . . , e
n−1
, e

n
) are orthogonal and span the same subspace. Then we can write
e

n
= a
1
e
1
+ . . . + a
n
e
n
. Now we have ¸e

n
, e
i
) = 0 for all i < n, so that a
i
= 0 for
i < n. Thus we have e

n
= a
n
e
n
and from |e

n
| = 1 we have a
n
= ±1.
Let (e
1
, . . . , e
n
) be the orthonormal base produced from (v
1
, . . . , v
n
) by Gram-Schmidt.
Then if (e

1
, . . . , e

n
) satisﬁes the hypothesis from the problem we have by the previous
paragraph that e

i
= ±e
i
. Thus we have 2
n
possible such orthonormal lists.
13. Extend (e
1
, . . . , e
m
) to a basis (e
1
, . . . , e
n
) of V . By theorem 6.17
v = ¸v, e
1
) e
1
+ . . . +¸v, e
n
) e
n
,
so that
|v| = | ¸v, e
1
) e
1
+ . . . +¸v, e
m
) e
n
|
= | ¸v, e
1
) e
1
| + . . . +| ¸v, e
m
) e
n
|
= [ ¸v, e
1
) [ + . . . +[ ¸v, e
n
) [.
Thus we have |v| = [ ¸v, e
1
) [ + . . . +[ ¸v, e
m
) [ if and only if ¸v, e
i
) = 0 for i > m i.e.
v ∈ span(e
1
, . . . , e
m
).
15
14. It’s easy to see that the diﬀerentiation operator has a upper-triangular matrix in the
orthonormal basis calculated in exercise 10.
15. This follows directly from V = U ⊕U

.
16. This follows directly from the previous exercise.
17. From exercise 5.21 we have that V = range P ⊕ null P. Let U := range P, then by
assumption null P ⊂ U

. From exercise 15, we have dimU = dimV − dimU =
dimnull P, so that null P = U

. An arbitrary v ∈ V can be written as v = Pv +(v −
Pv). From P
2
= P we have that P(v − Pv) = 0, so v − Pv ∈ null P = U

. Hence
the decomposition v = Pv + (v −Pv) is the unique decomposition in U ⊕U

. Now
Pv = P(Pv +(v −Pv)) = Pv, so that P is the identity on U. By deﬁnition P = P
U
.
18. Let u ∈ range P, then u = Pv for some v ∈ V hence Pu = P
2
v = Pv = u, so P is
the identity on range P. Let w ∈ null P. Then for a ∈ F
|u|
2
= |P(u + aw)|
2
≤ |u + aw|
2
.
By exercise 2 we have ¸u, w) = 0. Thus null P ⊂ (range P)

and from the dimension
equality null P = (range P)

. Hence the claim follows.
19. If TP
U
= P
U
TP
U
clearly U is invariant. Now assume that U is invariant. Then for
u ∈ U we have P
U
TP
U
u = P
U
Tu = Tu = TP
U
u.
20. Let u ∈ U, then we have Tu = TP
U
u = P
U
Tu ∈ U. Thus U is invariant. Then let
w ∈ U

. Now we can write Tw = u+u

where u ∈ U and u

∈ U

. Now P
U
Tw = u,
but u = P
U
Tw = TP
U
w = T0 = 0, so Tw ∈ U

. Thus U

is also invariant.
21. First we need to ﬁnd an orthonormal basis for U. With Gram-Schmidt we get e
1
=
(1/

2, 1/

2, 0, 0), e
2
= (0, 0, 1/

5, 2/

5). Let U = span(e
1
, e
2
). Then we have
u = P
U
(1, 2, 3, 4) = ¸(1, 2, 3, 4), e
1
) e
1
+¸(1, 2, 3, 4), e
2
) e
2
= (3/2, 3/2, 11/5, 22/5).
22. If p(0) = 0 and p

(0) = 0, then p(x) = ax
2
+ bx
3
. Thus we want to ﬁnd the
projection of 2 + 3x to the subspace U := span(x
2
, x
3
). With Gram-Schmidt we get
the orthonormal basis (

3x
2
,
_
420/11(x
3

1
2
x
2
)). We get
P
U
(2+3x) =
_
2 + 3x,

3x
2
_

3x
2
+
_
2 + 3x,
_
420/11(x
3

1
2
x
2
)
_
_
420/11(x
3

1
2
x
2
).
Here
_
2 + 3x,

3x
2
_
=
17
12

3,
_
2 + 3x,
_
420/11(x
3

1
2
x
2
)
_
=
47
660

1155,
so that
P
U
(2 + 3x) = −
71
22
x
2
+
329
22
x
3
.
16
23. There’s is nothing special with this exercise, compared to the previous two, except
that it takes ages to calculate.
24. We see that the map T : T
2
(R) → R deﬁned by p → p(1/2) is linear. From exercise
10 we get that (1,

3(2x − 1),

5(6x
2
− 6x + 1) is an orthonormal basis for T
2
(R).
From theorem 6.45 we see that
q(x) = 1 +

3(2 1/2 −1)

3(2x −1) +

5(6 (1/2)
2
−6 1/2 + 1)

5(6x
2
−6x + 1)
= −3/2 + 15x −15x
2
.
25. The map T : T
2
(R) → R deﬁned by p →
_
1
0
p(x) cos(πx)dx is clearly linear. Again
we have the orthonormal basis (1,

3(2x −1),

5(6x
2
−6x + 1), so that
q(x) =
_
1
0
cos(πx)dx +
_
1
0

3(2x −1) cos(πx)dx

3(2x −1)
+
_
1
0

5(6x
2
−6x + 1) cos(πx)dx

5(6x
2
−6x + 1)
= 0 −
4

3
π
2

3(2x −1) + 0 = −
12
π
2
(2x −1).
26. Choose an orthonormal basis (e
1
, . . . , e
n
) for V and let F have the usual basis. Then
by proposition 6.47 we have
/(T, (e
1
, . . . , e
n
)) = [¸e
1
, v) ¸e
n
, v)] ,
so that
/(T

, (e
1
, . . . , e
n
)) =
_
¸
_
¸e
1
, v)
.
.
.
¸e
n
, v)
_
¸
_
and ﬁnally T

a = (a¸e
1
, v), . . . , a¸e
n
, v)).
27. with the usual basis for F
n
we get
/(T) =
_
¸
¸
¸
¸
_
0 1 0
.
.
.
.
.
.
.
.
. 1
0 0
_
¸
¸
¸
¸
_
so by proposition 6.47
/(T

) =
_
¸
¸
¸
¸
_
0 0
1
.
.
.
.
.
.
.
.
.
0 1 0
_
¸
¸
¸
¸
_
.
17
Clearly T

(z
1
, . . . , z
n
) = (z
2
, . . . , z
n
, 0).
28. By additivity and conjugate homogeneity we have (T −λI)

= T

−λI. From exercise
31 we see that T

−λI is non-injective if and only if T −λI is non-injective. That is
λ is an eigenvalue of T

if and only if λ is an eigenvalue of T.
29. As (U

)

= U and (T

)

= T it’s enough to show only one of the implications. Let
U be invariant under T and choose w ∈ U

. Now we can write T

w = u + v, where
u ∈ U and v ∈ U

, so that
0 = ¸Tu, w) = ¸u, T

w) = ¸u, u + v) = ¸u, u) = |u|
2
.
Thus u = 0 which completes the proof.
30. As (T

)

= T it’s suﬃcient to prove the ﬁrst part. Assume that T is injective, then
by exercise 31 we have
dimV = dimrange T = dimrange T

,
but range T is a subspace of V , so that range T = V .
31. From proposition 6.46 and exercise 15 we get
dimnull T = dim(range T)

= dimW −dimrange T
= dimnull T + dimW −dimV.
For the second part we have
dimrangeT

= dimW −dimnullT

= dimV −dimnull T
= dimrange T,
where the second equality follows from the ﬁrst part.
32. Let T be the operator induced by A. The columns are the images of the basis vectors
under T, so the generate range T. Hence the dimension of the span of the column
vectors equal dimrange T. By proposition 6.47 the span of the row vectors equals
range T

. By the previous exercise dimrange T = dimrange T

, so the claim follows.
7 Operators on Inner-Product Spaces
1. For the ﬁrst part, let p(x) = x and q(x) = 1. Then clearly 1/2 = ¸Tp, q) , = ¸p, Tq) =
¸p, 0) = 0. For the second part it’s not a contradiction since the basis is not orthog-
onal.
18
2. Choose the standard basis for R
2
. Let T and S the the operators deﬁned by the
matrices
_
0 1
1 0
_
,
_
0 0
0 1
_
.
Then T and S as clearly self-adjoint, but TS has the matrix
_
0 1
0 0
_
.
3. If T and S are self-adjoint, then (aT +bS)

= (aT)

+(bS)

= aT +bS for any a, b ∈ R,
so self-adjoint operators form a subspace. The identity operator I is self-adjoint, but
clearly (iI)

= −iI, so iI is not self-adjoint.
4. Assume that P is self-adjoint. Then from proposition 6.46 we have null P = null P

=
(range P)

. Now P is clearly a projection to range P. Then assume that P is a
projection. Let (u
1
, . . . , u
n
) an orthogonal basis for range P and extend it to a basis
of V . Then clearly P has a diagonal matrix with respect to this basis, so P is
5. Choose the operators corresponding to the matrices
A =
_
0 −1
1 0
_
, B =
_
1 1
1 0
_
.
Then we easily see that A

A = AA

and B

B = BB

. However, A + B doesn’t
deﬁne a normal operator which is easy to check.
6. From proposition 7.6 we get null T = null T

. It follows from proposition 6.46 that
range T = (null T

)

= (null T)

= range T

.
7. Clearly null T ⊂ null T
k
. Let v ∈ null T
k
. Then we have
_
T

T
k−1
v, T

T
k−1
v
_
=
_
T

T
k
v, T
k−1
v
_
= 0,
so that T

T
k−1
v = 0. Now
_
T
k−1
v, T
k−1
v
_
=
_
T

T
k−1
v, T
k−2
v
_
= 0,
so that v ∈ null T
k−1
. Thus null T
k
⊂ null T
k−1
and continuing we get null T
k

null T.
Now let u ∈ range T
k
, then we can ﬁnd a v ∈ V such that u = T
k
v = T(T
k−1
v), so
that range T
k
⊂ range T. From the ﬁrst part we get dimrange T
k
= dimrange T, so
range T
k
= range T.
19
8. The vectors u = (1, 2, 3) and v = (2, 5, 7) are both eigenvectors corresponding to
diﬀerent eigenvalues. A self-adjoint operator is normal, so by corollary 7.8 if T is
normal that would imply orthogonality of u and v. Clearly ¸u, v) ,= 0, so T can’t be
9. If T is normal, we can choose a basis of V consisting of eigenvectors of T. Let A be
the matrix of T corresponding to the basis. Now T is self-adjoint if and only if the
conjugate transpose of A equals A that is the eigenvalues of T are real.
10. For any v ∈ V we get 0 = T
9
v −T
8
v = T
8
(Tv −v). Thus Tv −v ∈ null T
8
= null T
by the normality of T. Hence T(Tv −v) = T
2
v −Tv = 0, so T
2
= T. By the spectral
theorem we can choose a basis of eigenvectors for T such that T has a diagonal matrix
with λ
1
, . . . , λ
n
on the diagonal. Now T
2
has the diagonal matrix with λ
2
1
, . . . , λ
2
n
on
the diagonal and from T
2
= T we must have λ
2
i
= λ
i
for all i. Hence λ
i
= 0 or λ
i
= 1,
so the matrix of T equals its conjugate transpose. Hence T is self-adjoint.
11. By the spectral theorem we can choose a basis of T such that the matrix of T
corresponding to the basis is a diagonal matrix. Let λ
1
, . . . , λ
n
be the diagonal
elements. Let S be the operator corresponding to the diagonal matrix having the
elements

λ
1
, . . . ,

λ
n
on the diagonal. Clearly S
2
= T.
12. Let T the operator corresponding to the matrix
_
0 −1
1 0
_
. Then
T
2
+ I = 0.
13. By the spectral theorem we can choose a basis consisting of eigenvectors of T. Then
T has a diagonal matrix with respect to the basis. Let λ
1
, . . . , λ
n
be the diago-
nal elements. Let S be the operator corresponding to the diagonal matrix having
3

λ
1
, . . . ,
3

λ
n
on the diagonal. Then clearly S
3
= T.
14. By the spectral theorem we can choose a basis (v
1
, . . . , v
n
) consisting of eigenvectors
of T. Let v ∈ V be such that |v| = 1. If v = a
1
v
1
+ . . . + a
n
v
n
this implies that

n
i=1
[a
i
[ = 1. Assume that |Tv −λv| < ε. If [λ
i
−λ[ ≥ ε for all i, then
|Tv −λv| = |
n

i=1
(a
i
Tv
i
−λv
i
)| = |
n

i=1
(a
i
λ
i
v
i
−λv
i
)|
= |
n

i=1
a
i

i
−λ)v
i
| =
n

i=1
[a
i
[[λ
i
−λ[|v
1
|
=
n

i=1
[a
i
[[λ
i
−λ[ ≥ ε
n

i=1
[a
i
[ = ε,
which is a contradiction. Thus we can ﬁnd an eigenvalue λ
i
such that [λ −λ
i
[ < ε.
20
15. If such an inner-product exists we have a basis consisting of eigenvectors by the
spectral theorem. Assume then that (v
1
, . . . , v
n
) is basis such that the matrix of T
is a matrix consisting of eigenvectors of T. Deﬁne an inner-product by
¸v
i
, v
j
) =
_
0, i ,= j
1, i = j
.
Extend this to an inner-product by bilinearity and homogeneity. Then we have an
inner-product and clearly T is self-adjoint as (v
1
, . . . , v
n
) is an orthonormal basis of
U.
16. Let T ∈ L(R
2
) be the operator T(a, b) = (a, a + b). Then span((1, 0)) is invariant
under T, but it’s orthogonal complement span((0, 1)) is clearly not.
17. Let T, S be two positive operators. Then they are self-adjoint and (S + T)

=
S

+T

= S +T, so S +T is self-adjoint. Also ¸(S + T)v, v) = ¸Sv, v) +¸Tv, v) ≥ 0.
Thus, S + T is positive.
18. Clearly T
k
is self-adjoint for every positive integer. For k = 2 we have
¸
T
2
v, v
_
=
¸Tv, Tv) ≥ 0. Assume that the result is true for all positive k < n and n ≥ 2. Then
¸T
n
v, v) =
¸
T
n−2
Tv, Tv
_
≥ 0,
by hypothesis and self-adjointness. Hence the result follows by induction.
19. Clearly ¸Tv, v) > 0 for all v ∈ V ¸ ¦0¦ implies that T is injective hence invertible.
So assume that T is invertible. Since T is self-adjoint we can choose an orthonormal
basis (v
1
, . . . , v
n
) of eigenvectors such that T has a diagonal matrix with respect to
the basis. Let the elements on the diagonal be λ
1
, . . . , λ
n
and by assumption λ
i
> 0
for all i. Let v = a
1
v
1
+ . . . + a
n
v
n
∈ V ¸ ¦0¦, so that
¸Tv, v) = [a
1
[
2
λ
1
+ . . . +[a
n
[
2
λ
n
> 0.
20.
_
sinθ cos θ
cos θ −sinθ
_
is a square root for every θ ∈ R, which shows that it has inﬁnitely
many square roots.
21. Let S ∈ L(R
2
) be deﬁned by S(a, b) = (a + b, 0). Then |S(1, 0)| = |(1, 0)| = 1 and
|S(0, 1)| = 1, but S is clearly not an isometry.
22. R
3
is an odd-dimensional real vector space. Hence S has an eigenvalue λ and a
corresponding eigenvector v. Since |Sv| = [λ[|v| = |v| we have λ
2
= 1. Hence
S
2
v = λ
2
v = v.
21
23. The matrices corresponding to T and T

are
/(T) =
_
_
0 0 1
2 0 0
0 3 0
_
_
, /(T

) =
_
_
0 2 0
0 0 3
1 0 0
_
_
.
Thus we get
/(T

T) =
_
_
4 0 0
0 9 0
0 0 1
_
_
,
so that T

T is the operator (z
1
, z
2
, z
3
) → (4z
1
, 9z
2
, z
3
). Hence

T

T(z
1
, z
2
, z
3
) =
(2z
1
, 3z
2
, z
3
). Now we just need to permute the indices which corresponds to the
isometry S(z
1
, z
2
, z
3
) = (z
3
, z
1
, z
2
).
24. Clearly T

T is positive, so by proposition 7.26 it has a unique positive square root.
Thus it’s suﬃcient to show that R
2
= T

T. Now we have T

= (SR)

= R

S

= RS

by self-adjointness of R. Thus R
2
= RIR = RS

SR = T

T.
25. Assume that T is invertible. Then

T

T must be invertible (polar decomposition).
Hence S = T

T

T
−1
, so S is uniquely determined. Now assume that T is not
invertible. Then range

T

T is not invertible (hence not surjective).
Now assume that T is not invertible, so that

T

T is not invertible. By the spectral
theorem we can choose a basis (v
1
, . . . , v
n
) of eigenvectors of

T

T and we can
assume that v
1
corresponds to the eigenvalue 0. Now let T = S

T

T. Deﬁne U
by Uv
1
= −Sv
1
, Uv
i
= Sv
i
, i > 1. Clearly U ,= S and T = U

T

T. Now choose
u, v ∈ V . Then it’s easy to verify that ¸Uu, Uv) = ¸Su, Sv) = ¸u, v), so U is an
isometry.
26. Choose a basis of eigenvectors for T such that T has a diagonal matrix with eigenval-
ues λ
1
, . . . , λ
n
on the diagonal. Now T

T = T
2
corresponds to the diagonal matrix
with λ
2
1
, . . . , λ
2
n
on the diagonal and the square root

T

T corresponds to the diag-
onal matrix having [λ
1
[, . . . , [λ
n
[ on the diagonal. These are the singular values, so
the claim follows.
27. Let T ∈ L(R
2
) be the map T(a, b) = (0, a). Then from the matrix representation
we see that T

T(a, 0) = (a, 0), hence

T

T = T

T and 1 is clearly a singular value.
However, T
2
= 0, so
_
(T
2
)

T
2
= 0, so 1
2
= 1 is not a singular value of T
2
.
28. The composition of two bijections is a bijection. If T=S

T

T, then since S is an
isometry we have that T is bijective hence invertible if and only if

T

T is bijective.
But

T

T is injective hence bijective if and only if 0 is not an eigenvalue. Thus T is
bijective if and only if 0 is not a singular value.
22
29. From the polar decomposition theorem we see that dimrange T = dimrange

T

T.
Since

T

T is self-adjoint we can choose a basis of eigenvectors of

T

T such that
the matrix of

T

T is a diagonal matrix. Clearly dimrange

T

T equals the number
of non-zero elements on the diagonal i.e. the number of non-zero singular values.
30. If S is an isometry, then clearly

S

S = I, so all singular values are 1. Now assume
that all singular values are 1. Then

S

S is a self-adjoint (positive) operator with
all eigenvalues equal to 1. By the spectral theorem we can choose a basis of V such
that

S

S has a diagonal matrix. As all eigenvalues are one, this means that the
matrix of

S

S is the identity matrix. Thus

S

S = I, so S is an isometry.
31. Let T
1
= S
1
_
T

1
T
1
and T
2
= S
2
_
T

2
T
2
. Assume that T
1
and T
2
have the same
singular values s
1
, . . . , s
n
. Then we can choose bases of eigenvectors of
_
T

1
T
1
and
_
T

2
T
2
such that they have the same matrix. Let these bases be (v
1
, . . . , v
n
) and
(w
1
, . . . , w
n
). Let S the operator deﬁned by S(w
i
) = v
i
, then clearly S is an isometry
and
_
T

2
T
2
= S
−1
_
T

1
T
1
S. Thus we get T
2
= S
2
_
T

2
T
2
= S
2
S
−1
_
T

1
T
1
S. Writing
S
3
= S
2
S
−1
S
−1
1
, which is clearly an isometry, we get
T
2
= S
3
S
1
_
T

1
T
1
S = S
3
T
1
S.
Now let T
2
= S
1
T
1
S
2
. Then we have T

2
T
2
= (S
1
T
1
S
2
)

(S
1
T
1
S
2
) = S

1
T

1
T
1
S
1
=
S
−1
1
T

1
T
1
S
1
. Let λ be an eigenvalue of T

1
T
1
and v a corresponding eigenvector.
Because S
1
is bijective we have a u ∈ V such that v = S
1
u. This gives us
T

2
T
2
u = S
−1
1
T

1
T
1
S
1
u = S
−1
1
T

1
T
1
v = S
−1
1
λv = λu,
so that λ is an eigenvalue of T

2
T
2
. Let (v
1
, . . . , v
n
) be an orthonormal basis of eigen-
vectors of T

1
T
1
, then (S
−1
1
v
1
, . . . , S
−1
v
n
) is an orthonormal basis for V of eigenvectors
of T

2
T
2
. Hence T

2
T
2
and T

1
T
1
have the same eigenvalues. The singular values are
simply the positive square roots of these eigenvalues, so the claim follows.
32. For the ﬁrst part, denote by S the map deﬁned by the formula. Set
A := /(S, (f
1
, . . . , f
n
), (e
1
, . . . , e
n
)), B := /(T, (e
1
, . . . , e
n
), (f
1
, . . . , f
n
)).
Clearly A = B and is a diagonal matrix with s
1
, . . . , s
n
on the diagonal. All the s
i
are positive real numberse, so that B

= B = A. It follows from proposition 6.47
that S = T

.
For the second part observe that a linear map S deﬁned by the formula maps f
i
to s
−1
i
e
i
which is mapped by T to f
i
. Thus TS = I. The claim now follows from
exercise 3.23.
23
33. By deﬁnition

T

T is positive. Thus all singular values of T are positive. From the
singular value decomposition theorem we can choose bases of V such that
|Tv| = |
n

i=1
s
i
¸v, e
i
) f
i
| =
n

i=1
s
i
[ ¸v, e
i
) [
since all s
i
are positive. Now clearly
ˆ s|v| = ˆ s
n

i=1
[ ¸v, e
i
) [ ≤
n

i=1
s
i
[ ¸v, e
i
) [ = |Tv|
and similarly for s|v|.
34. From the triangle equality and the previous exercise we get
|(T

+ T

)v| ≤ |T

v| +|T

v| ≤ (s

+ s

)|v|.
Now let v be the eigenvector of
_
(T

+ T

)

(T

+ T

) corresponding to the singular
value s. Let T + T

= S
_
(T

+ T

)

(T

+ T

), so that
|(T

+ T

)v| = |S
_
(T

+ T

)

(T

+ T

)v| = |Ssv| = s|v| ≤ (s

+ s

)|v|,
so that s ≤ s

+ s

.
8 Operators on Complex Vector Spaces
1. T is not injective, so 0 is an eigenvalue of T. We see that T
2
= 0, so that any v ∈ V
is a generalized eigenvector corresponding to the eigenvalue 0.
2. On page 78 it’s shown that ±i are the eigenvalues of T. We have dimnull(T −iI)
2
+
dimnull(T + iI)
2
= dimC
2
= 2, so that dimnull(T − iI)
2
= dimnull(T + iI)
2
= 1.
Because (1, −i) is an eigenvector corresponding to the eigenvalue i and (1, i) is an
eigenvector corresponding to the eigenvalue −i. The set of all generalized eigenvectors
are simply the spans of these corresponding eigenvectors.
3. Let

m−1
i=0
a
i
T
i
v = 0, then 0 = T
m−1
(

m−1
i=0
a
i
T
i
v = a
0
T
m−1
v, so that a
0
= 0. Ap-
plying repeatedly T
m−i
for i = 2, . . . we see that a
i
= 0 for all i. Hence (v, . . . , T
m−1
v)
is linearly independent.
4. We see that T
3
= 0, but T
2
,= 0. Assume that S is a square root of T, then S
is nilpotent, so by corollary 8.8 we have S
dimV
= S
3
= 0, so that 0 = S
4
= T
2
.
5. If ST is nilpotent, then ∃n ∈ N such that (ST)
n
= 0, so (TS)
n+1
= T(ST)
n
S = 0.
24
6. Assume that λ ,= 0 is an eigenvalue of N with corresponding eigenvector v. Then
N
dimV
v = λ
dimV
v ,= 0. This contradicts corollary 8.8.
7. By lemma 8.26 we can ﬁnd a basis of V such that the matrix of N is upper-triangular
with zeroes on the diagonal. Now the matrix of N

is the conjugate transpose. Since
N

= N, the matrix of N must be zero, so the claim follows.
8. If N
dimV −1
,= N
dimV
, then dimnull N
i−1
< dimnull N
i
for all i ≤ dimV by propo-
sition 8.5. By corollary dimnull N
dimV
= dimV , so the claim clearly follows.
9. range T
m
= range T
m+1
implies dimnull T
m
= dimnull T
m+1
i.e. null T
m
= null T
m+1
.
From proposition 8.5 we get dimnull T
m
= dimnull T
m+k
for all k ≥ 1. Again this
implies that dimrange T
m
= dimrange T
m+k
for all k ≥ 1, so the claim follows.
10. Let T be the operator deﬁned in exercise 1. Clearly null T ∩ range T ,= ¦0¦, so the
claim is false.
11. We have dimV = dimnull T
n
+dimrange T
n
, so it’s suﬃcient to prove that null T
n

range T
n
= ¦0¦. Let v ∈ null T
n
∩ range T
n
. Then we can ﬁnd a u ∈ V such that
T
n
u = v. From 0 = Tv = T
n+1
u we see that u ∈ null T
n+1
= null T
n
which implies
that v = T
n
u = 0.
12. From theorem 8.23 we have V = null T
dimV
, so that T
dimV
= 0. Then let T ∈ L(R
3
)
be the operator T(a, b, c) = (−b, a, 0). Then clearly 0 is the only eigenvalue, but T is
not nilpotent.
13. From null T
n−2
,= null T
n−1
and from proposition 8.5 we see that dimnull T
n
≥ n−1.
Assume that T has three diﬀerent eigenvalues 0, λ
1
, λ
2
. Then dimnull (T −λ
i
I)
n
≥ 1,
so from theorem 8.23
n ≥ dimnull T
n
+dimnull (T −λ
1
I)
n
+dimnull (T −λ
2
I)
n
≥ n −1 +1 +1 = n +1,
which is impossible, so T has at most two diﬀerent eigenvalues.
14. Let T ∈ L(C
4
) be deﬁned by T(a, b, c, d) = (7a, 7b, 8c, 8d) from the matrix of T it’s
easy to see that the characteristic polynomial is (z −7)
2
(z −8)
2
.
15. Let d
1
be the multiplicity of the eigenvalue 5 and d
2
the multiplicity of the eigenvalue
6. Then d
1
, d
2
≥ 1 and d
1
+ d
2
= n. It follows that d
1
, d
2
≤ n − 1, so that
(z − 5)
d
1
(z − 6)
d
2
divides (z − 5)
n−1
(z − 6)
n−1
. By the Cayley-Hamilton theorem
(T −5I)
d
1
(T −6I)
d
2
= 0, so that (T −5I)
n−1
(T −6I)
n−1
= 0.
16. If every generalized eigenvector is an eigenvector, then by theorem 8.23 T has a basis
of eigenvectors. If there’s a generalized eigenvector that is not an eigenvector, then
we have an eigenvalue λ such that dimnull(T −λI)
dimV
,= dimnull(T −λI). Thus if
25
λ
1
, . . . , λ
m
are the eigenvalues of T, then

m
i=1
dimnull(T − λ
i
I) < dimV , so there
doesn’t exists a basis of eigenvectors.
17. By lemma 8.26, choose a basis (v
1
, . . . , v
n
) such that N has an upper-triangular
matrix. Apply Gram-Schmidt orthogonalization to the basis. Then Ne
1
= 0 and
assume that Ne
i
∈ span(e
1
, . . . , e
i
), so that
Ne
i+1
= N
_
v
i+1
−¸v
i+1
, e
1
) e
1
−. . . −¸v
i+1
, e
i
) e
i
|v
i+1
−¸v
i+1
, e
1
) e
1
−. . . −¸v
i+1
, e
i
) e
i
|
_
∈ span(e
1
, . . . , e
i+1
).
Thus N has an upper-triangular matrix in the basis (e
1
, . . . , e
n
).
18. Continuing in the proof of lemma 8.30 up to j = 4 we see that a
4
= −5/128. So that

I + N = I +
1
2
N −
1
8
N
2
+
1
16
N
3

5
128
N
4
19. Just replace the Taylor-polynomial in lemma 8.30 with the Taylor polynomial of
3

1 + x and copy the proof of the lemma and theorem 8.32.
20. Let p(x) =

m
i=0
a
i
x
i
be the minimal polynomial of T. If a
0
,= 0, then p(x) =
x

m
i=1
a
i
x
i−1
. Since T is invertible we must have

m
i=1
a
i
T
i−1
the minimality of p. Hence a
0
,= 0, so solving for I in p(T) we get
−a
−1
0
(a
1
+ . . . + a
m
T
m−1
)T = I.
Hence setting q(x) = −a
−1
0
(a
1
+ . . . + a
m
x
m−1
) we have q(T) = T
−1
.
21. The operator deﬁned by the matrix
_
_
0 0 1
0 0 0
0 0 0
_
_
is clearly an example.
22. Choose the matrix
A =
_
¸
¸
_
1 0 0 0
0 1 1 0
0 0 1 0
0 0 0 0
_
¸
¸
_
.
It’s easy to see that A(A−I)
2
= 0. Thus, the minimal polynomial divides z(z −1)
2
.
However, the whole polynomial is the only factor that annihilates A, so z(z −1)
2
is
the minimal polynomial.
26
23. Let λ
1
, . . . , λ
m
be the eigenvalues of T with multiplicity d
1
, . . . , d
m
. Assume that V
has a basis of eigenvectors (v
1
, . . . , v
n
). Let p(z) =

m
i=1
(z −λ
i
). Then we get
p(T)v
i
=
_
_

j=i
(T −λ
j
I)
_
_
(T −λ
i
I)v
i
= 0,
so that the minimal polynomial divides p and hence has no double root.
Now assume that the minimal polynomial has no double roots. Let the minimal
polynomial be p and let (v
1
, . . . , v
n
) be a Jordan-basis of T. Let A be the largest
Jordan-block of T. Now clearly p(A) = 0. However, by exercise 29, the minimal
polynomial of (A − λI) is z
m+1
where m is the length of the longest consecutive
string of 1’s that appear just above the diagonal, so the minimal polynomial of A
is (z − λ)
m+1
. Hence m = 0, so that A is a 1 1 matrix and the basis vector
corresponding to A is an eigenvalue. Since A was the largest Jordan-block it follows
that all basis vectors are eigenvalues (corresponding to a 1 1 matrix).
24. If V is a complex vector space, then V has a basis of eigenvectors. Now let λ
1
, . . . , λ
m
be the distinct eigenvalues. Now clearly p =

m
i=1
(z −λ
i
) annihilates T, so that the
minimal polynomial being a factor of p doesn’t have a double root.
Now assume that V is a real vector space. By theorem 7.25 we can ﬁnd a basis of T
such that T has a block diagonal matrix where each block is a 1 1 or 2 2 matrix.
Let λ
1
, . . . , λ
m
be the distinct eigenvalues corresponding to the 1 1 blocks. Then
p(z) =

m
i=1
(z − λ
i
) annihilates all but the 2 2 blocks of the matrix. Now it’s
suﬃcient to show that each 2 2 block is annihilated by a polynomial which doesn’t
have real roots.
By theorem 7.25 we can choose the 2 2 blocks to be of the form
_
a −b
b a
_
where b > 0 and it’s easy to see that
_
a −b
b a
_
−2a
_
a −b
b a
_
+ (a
2
+ b
2
)
_
1 0
0 1
_
= 0.
Clearly this polynomial has a negative discriminant, so the claim follows.
25. Let q be the minimal polynomial. Write q = sp + r, where deg r < deg p, so that
0 = q(T)v = s(T)p(T)v +r(T)v = r(T)v. Assuming r ,= 0, we can multiply the both
sides with the inverse of the highest coeﬃcient of r yielding a monic polynomial r
2
of degree less than p such that r
2
(T)v = 0 contradicting the minimality of p. Hence
r = 0 and p divides q.
27
26. It’s easy to see that no proper factor of z(z −1)
2
(z −3) annihilates the matrix
A =
_
¸
¸
_
3 1 1 1
0 1 1 0
0 0 1 0
0 0 0 0
_
¸
¸
_
,
so the it’s minimal polynomial is z(z − 1)
2
(z − 3) which by deﬁnition is also the
characteristic polynomial.
27. It’s easy to see that no proper factor of z(z −1)(z −3) annihilates the matrix
A =
_
¸
¸
_
3 1 1 1
0 1 1 0
0 0 1 0
0 0 0 0
_
¸
¸
_
,
but A(A−I)(A−3I) = 0, so the claim follows.
28. We see that T(e
i
) = e
i+1
for i < n. Hence (e
1
, Te
1
, . . . , T
n−1
e
1
) = (e
1
, . . . , e
n
) is
linearly independent. Thus for any non-zero polynomial p of degree less than n we
have p(e
i
) ,= 0. Hence the minimal polynomial has degree n, so that the minimal
polynomial equals the characteristic polynomial.
Now from the matrix of T we see that T
n
(e
i
) = −a
0
−a
1
T(e
1
) −. . . −a
n−1
T
n−1
. Set
p(z) = a
0
+a
1
z +. . . +a
n−1
z
n−1
+z
n
. Now p(T)(e
1
) = 0, so by exercise 25 p divides
the minimal polynomial. However, the minimal polynomial is monic and has degree
n, so p is the minimal polynomial.
29. The biggest Jordan-block of N is of the form
_
¸
¸
¸
¸
¸
_
0 1
0 1
.
.
.
.
.
.
0 1
0
_
¸
¸
¸
¸
¸
_
.
Now clearly N
m+1
= 0, so the minimal polynomial divides z
m+1
. Let v
i
, . . . , v
i+m
be
the basis elements corresponding to the biggest Jordan-block. Then N
m
e
i+m
= e
i
,=
0, so the minimal polynomial is z
m+1
.
30. Assume that V can’t be decomposed into two proper subspaces. Then T has only
one eigenvalue. If there is more than one Jordan-block, then Let (v
1
, . . . , v
m
) be
the vectors corresponding to all but the last Jordan-block and (v
m+1
, . . . , v
n
) the
28
vectors corresponding to the last Jordan-block. Then clearly V = span(v
1
, . . . , v
m
) ⊕
span(v
m+1
, . . . , v
n
). Thus, we have only one Jordan-block and T − λI is nilpotent
and has minimal polynomial z
dimV
by the previous exercise. Hence T has minimal
polynomial (z −λ)
dimV
.
Now assume that T has minimal polynomial (z −λ)
dimV
. If V = U ⊕W, let p
1
be the
minimal polynomial of T
|U
and p
2
the minimal polynomial of T
|W
. Now (p
1
p
2
)(T) =
0, so that (z − λ)
dimV
divides p
1
p
2
. Hence deg p
1
p
2
= deg p
1
+ deg p
2
≥ dimV , but
deg p
1
+deg p
2
≤ dimU +dimW = dimV . Thus deg p
1
p
2
= dimV . This means that
p
1
(z) = (z − λ)
dimU
and p
2
(z) = (z − λ)
dimW
. Now if n = max¦dimU, dimW¦ we
have (T
|U
− λI)
n
= (T
|W
− λI)
n
= 0. This means that (T − λI)
n
the fact that (z −λ)
dimV
is the minimal polynomial.
31. Reversing the Jordan-basis simply reverses the order of the Jordan-blocks and each
block needs to be replaced by its transpose.
9 Operators on Real Vector Spaces
1. Let a be the value in the upper-left corner. Then the matrix must have the form
_
a 1 −a
1 −a a
_
and it’s easy to see that
_
1
1
_
is an eigenvector corresponding to the eigenvalue 1.
2. The characteristic polynomial of the matrix is p(x) = (x − a)(x − d) − bc. Now the
discriminant equals (a + d)
2
2
+ 4bc. Hence the characteristic
polynomial has a root if and only if (a − d)
2
+ 4bc ≥ 0. If p(x) = (x − λ
1
)(x − λ
2
),
then p(T) = 0 implies that either A − λ
1
I or A − λ
2
I is not invertible hence A has
an eigenvalue. If p doesn’t have a root then assuming that Av = λv we get
0 = p(A)v = p(λ)v,
so that v = 0. Hence A has no eigenvalues. The claim follows.
3. See the proof of the next problem, though in this special case the proof is trivial.
4. First let λ be an eigenvalue of A. Assume that λ is not an eigenvalue of A
1
, . . . , A
m
,
so that A
i
−λI is invertible for each i. Now let B be the matrix
_
¸
_
(A
1
−λI)
−1

.
.
.
0 (A
m
−λI)
−1
_
¸
_.
29
It’s now easy to verify that B(A−λI) is a diagonal matrix with all 1s on the diagonal,
hence it’s invertible. However, this is impossible, since A −λI is not. Thus, λ is an
eigenvalue of some A
i
.
Now let λ be an eigenvalue of A
i
. Let T be the operator corresponding to A − λI
in L(F
n
). Let A
i
correspond to the columns j, . . . , j + k and let (e
1
, . . . , e
n
) be the
standard basis. Let (a
1
, . . . , a
k
) be the eigenvector of A
i
corresponding to λ. Set v =
a
1
e
j
+. . . +a
k
e
j+k
. Then it’s easy to see that T maps the space span(e
1
, . . . , e
j−1
, v)
to the space span(e
1
, . . . , e
j−1
). Hence T is not injective, so that λ is an eigenvalue
of A.
5. The proof is identical to the argument used in the solution of exercise 2.
6. Let (v
1
, . . . , v
n
) be the basis with the respect to which T has the given matrix.
Applying Gram-Schmidt to the basis the matrix trivially has the same form.
7. Choose a basis (v
1
, . . . , v
n
) such that T has a matrix of the form in theorem 9.10.
Now if v
j
corresponds to the second vector in a pair corresponding to a 2 2 matrix
or corresponds to a 1 1 matrix, then span(v
1
, . . . , v
j
) is an invariant subscape. The
second posibility means that v
j
corresponds to the ﬁrst vector in a pair corresponding
to a 2 2 matrix, so that span(v
1
, . . . , v
j+1
) is an invariant subspace.
8. Assuming that such an operator existed we would have basis (v
1
, . . . , v
7
) such that
the matrix of T would have a matrix with eigenpair (1, 1)
dimnull(T
2
+ T + I)
dimV
2
= 7/2
times on the diagonal. This would contradict theorem 9.9 as 7/2 is not a whole
number. It follows that such an operator doesn’t exist.
9. The equation x
2
+ x + 1 = 0 has a solution in C. Let λ be a solution. Then the
operator corresponding to the matrix λI clearly is an example.
10. Let T be the operator in L(R
2k
) corresponding to the block diagonal matrix where
each block has characteristic polynomial x
2
+αx +β. Then by theorem 9.9 we have
k =
dimnull(T
2
+ αT + βI)
2k
2
,
so that dimnull(T
2
+ αT + βI)
2k
is even. However, the minimal polynomial of T
divides (T
2
+αT +βI)
2k
and has degree less than 2k, so that (T
2
+αT +βI)
k
= 0.
It follows that dimnull(T
2
+ αT + βI)
2k
= dimnull(T
2
+ αT + βI)
k
and the claim
follows.
30
11. We see that (α, β) is an eigenpair and from the nilpotency of T
2
+ αT + βI and
theorem 9.9 we have
dimnull(T
2
+ αT + βI)
dimV
2
= (dimV )/2
it follows that dimV is even. For the second part we know from the Cayley-Hamilton
theorem that the minimal polynomial p(x) of T has degree less than dimV . Thus we
have p(x) = (x
2
+ αx + β)
k
and from deg p ≤ dimV we get k ≤ (dimV )/2. Hence
(T + αT + βI)
(dimV )/2
= 0.
12. By theorem 9.9 we can choose a basis such that the matrix of T has the form of 9.10.
Now we have the matrices [5] and [7] at least once on the diagonal. Assuming that
T has an eigenpair we would have at least one 2 2 matrix on the diagonal. This is
impossible as the diagonal has only length 3.
13. By proposition 8.5 dimnull T
n−1
≥ n−1. Like in the previous exercise we know that
the matrix [0] is at least n−1 times on the diagonal and it’s impossible to ﬁt a 2 2
matrix in the only place left.
14. We see that A satisﬁes the polynomial p(z) = (z − a)(z − d) − bc no matter if F is
R or C. If F = C, then because p is monic, has degree 2 and annihilates A it follows
that p is the characteristic polynomial.
Assume then that F = R. Now if A has no eigenvalues, then p is the characteristic
polynomial by deﬁnition. Assume then that p(z) = (z − λ
1
)(z − λ
2
) which implies
that p(T) = (T −λ
1
I)(T −λ
2
I) = 0. If neither T −λ
1
I or T −λ
2
I is injective, then
by deﬁnition p is the minimal polynomial. Assume then that T − λ
2
I is invertible.
Then dimnull(T − λ
1
I) = 2, so that T − λ
1
I = 0. Hence we get c = b = 0 and
a = d = λ
1
. Hence p(z) = (z − λ
1
)
2
, so that p is the characteristic polynomial by
deﬁnition.
15. S is normal, so there’s an orthonormal basis of V such that S has a block diagonal
matrix with respect to the basis and each block is a 1 1 or 2 2 matrix and the
22 blocks have no eigenvalue (theorem 7.25). From S

S = I we see that each block
A
i
satisfy A

i
A
i
= I, so that they are isometries. Hence a 2 2 block is of the form
_
cos θ −sinθ
sinθ cos θ
_
.
We see that T
2
+αT +βI is not injective if and only if A
2
i
+αA
i
+βI is not injective
for some 22 matrix A
i
. Hence x
2
+αx+β must equal the characteristic polynomial
of some A
i
, so by the previous exercise it’s of the form
(x −cos θ)(z −cos θ) + sin
2
θ = x
2
−2xcos θ + cos
2
θ + sin
2
θ
so that β = cos
2
θ + sin
2
θ = 1.
31
10 Trace and Determinant
1. The map T → /(T, (v
1
, . . . , v
n
)) is bijective and satisﬁes ST → /(ST, (v
1
, . . . , v
n
)) =
/(S, (v
1
, . . . , v
n
))/(T, (v
1
, . . . , v
n
)). Hence
ST = I ⇔ /(ST, (v
1
, . . . , v
n
)) = /(S, (v
1
, . . . , v
n
))/(T, (v
1
, . . . , v
n
)) = I.
The claim now follows trivially.
2. Both matrices represent an operator in L(F
n
). The claim now follows from exercise
3.23.
3. Choose a basis (v
1
, . . . , v
n
) and let A = (a
ij
) be the matrix corresponding to the
basis. Then Tv
1
= a
11
v
1
+ . . . + a
n1
v
n
. Now (v
1
, 2v
2
, . . . , 2v
n
) is also a basis and we
have by our assumption Tv = a
11
v
1
+ 2(a
21
v
2
+ . . . + a
n1
v
n
). We thus get
a
21
v
2
+ . . . + a
n1
v
n
= 2(a
21
v
2
+ . . . + a
n1
v
n
),
which implies (a
21
v
2
+ . . . + a
n1
v
n
= 0. By linear independence we get a
21
= . . . =
a
n1
= 0, so that Tv
1
= a
11
v. The claim clearly follows.
4. Follows directly from the deﬁnition of /(T, (u
1
, . . . , u
n
), (v
1
, . . . , v
n
)).
5. Let (e
1
, . . . , e
n
) be the standard basis for C
n
. Then we can ﬁnd an operator T ∈
L(C
n
) such that /(T, (e
1
, . . . , e
n
)) = B. Now T has an upper-triangular matrix
corresponding to some basis (v
1
, . . . , v
n
) of V . Let A = /((v
1
, . . . , v
n
), (e
1
, . . . , e
n
)).
Then
A
−1
BA = /(T, (v
1
, . . . , v
n
))
which is upper-triangular. Clearly A is an invertible square matrix.
6. Let T be the operator corresponding to the matrix
_
0 −1
1 0
_
,
_
0 −1
1 0
_
2
=
_
−1 0
0 −1
_
.
By theorem 10.11 we have trace(T
2
) = −2 < 0.
7. Let (v
1
, . . . , v
n
) be a basis of eigenvectors of T and λ
1
, . . . , λ
n
be the corresponding
eigenvalues. Then T has the matrix
_
¸
_
λ
1
0
.
.
.
0 λ
n
_
¸
_.
Clearly trace(T
2
) = λ
2
1
+ . . . + λ
2
n
≥ 0 by theorem 10.11.
32
8. Extend v/|v| to an orthonormal basis (v/|v|, e
1
, . . . , e
n
) of V and let A = /(T, (v/|v|, e
1
, . . . , e
n
)).
Let a, a
1
, . . . , a
n
denote the diagonal elements of A. Then we have
a
i
= ¸Te
i
, e
i
) = ¸¸e
i
, v) w, e
i
) = ¸0, e
i
) = 0,
so that
trace(T) = a = ¸Tv/|v|, v/|v|) = ¸¸v/|v|, v) w, v/|v|) = ¸w, v) .
9. From exercise 5.21 we have V = null P ⊕range P. Now let v ∈ range P. Then v = Pu
for some u ∈ V . Hence Pv = P
2
w = Pw = v, so that P is the identity on range P.
Now choose a basis (v
1
, . . . , v
m
) for range P and extend it with a basis (u
1
, . . . , u
n
)
of null P to get a basis for V . Then clearly the matrix for P in this basis consists
of a diagonal matrix with 1s on part of the diagonal corresponding to the vectors
(v
1
, . . . , v
m
) and 0s on the rest of the diagonal. Hence trace P = dimrange P ≥ 0.
10. Let (v
1
, . . . , v
n
) be some basis of V and let A = /(T, (v
1
, . . . , v
n
)). Let a
1
, . . . , a
n
be
the diagonal elements of A. By proposition 6.47 T

has the matrix A

, so that the
diagonal elements of A

are a
1
, . . . , a
n
. By theorem 10.11 we have
trace(T

) = a
1
+ . . . + a
n
= a
1
+ . . . + a
n
= trace(T).
11. A positive operator is self-adjoint. By the spectral theorem we can ﬁnd a basis
(v
1
, . . . , v
n
) of eigenvectors of T. Then A = /(T, (v
1
, . . . , v
n
)) is a diagonal matrix
and by positivity of T all the diagonal elements a
1
, . . . , a
n
are positive. Hence a
1
+
. . . + a
n
= 0 implies a
1
= . . . = a
n
= 0, so that A = 0 and hence T = 0.
12. The trace of T is the sum of the eigenvalues. Hence λ − 48 + 24 = 51 − 40 + 1, so
that λ = 36.
13. Choose a basis of (v
1
, . . . , v
n
) of V and let A = /(T, (v
1
, . . . , v
n
)). Then the matrix
of cT is cA. Let a
1
, . . . , a
n
be the diagonal elements of A. Then we have
trace(cT) = ca
1
+ . . . + ca
n
= c(a
1
+ . . . + a
n
) = ctrace(T).
14. The example in exercise 6 shows that this is false. For another example take S =
T = I.
15. Choose a basis (v
1
, . . . , v
n
) for V and let A = /(T, (v
1
, . . . , v
n
)). It’s suﬃcient to
prove that A = 0. Now let a
i,j
be the element in row i, column j of A. Let B be
the matrix with 1 in row j column i and 0 elsewhere. Then it’s easy to see that in
BA the only non-zero diagonal element is a
i,j
. Thus trace(BA) = a
i,j
. Let S be the
operator corresponding to B. It follows that a
i,j
= trace(ST) = 0, so that A = 0.
33
16. Let T

Te
i
= a
1
e
1
+ . . . + a
n
e
n
. Then we have that a
i
= ¸T

Te
i
, e
i
) = |Te
i
|
2
by
the ortogonality of (e
1
, . . . , e
n
). Clearly a
i
is the ith diagonal element in the matrix
/(T

T, (e
1
, . . . , e
n
)), so that
trace(T

T) = |Te
1
|
2
+ . . . +|Te
n
|
2
.
The second assertion follows immediately.
17. Choose a basis (v
1
, . . . , v
n
) such that T has an upper triangular matrix B = (b
ij
)
with eigenvalues λ
1
, . . . , λ
n
on the diagonal. Now the ith element on the diagonal of
B

B is

n
j=1
b
ji
b
ji
=

n
j=1
[b
ji
[
2
, where [b
ii
[
2
= [λ
i
[
2
. Hence
n

i=1

i
[
2

n

i=1
n

j=1
[b
ji
[
2
= trace(B

B) = trace(A

A) =
n

k=1
n

j=1
[a
jk
[
2
.
18. Positivity and deﬁniteness follows from exercise 16 and additivity in the ﬁrst slot
from corollary 10.12. Homogeneity in the ﬁrst slot is exercise 13 and by exercise 10
¸S, T) = trace(ST

) = trace((ST

)

) = trace(TS

) = ¸T, S).
It follows that the formula deﬁnes an inner-product.
19. We have 0 ≤ ¸Tv, Tv) − ¸T

v, T

v) = ¸(T

T −TT

)v, v). Hence T

T − TT

is a
positive operator (it’s clearly self-adjoint), but trace(T

T − TT

) = 0, so all the
eigenvalues are 0. Hence T

T −TT

= 0 by the spectral theorem, so T is normal.
20. Let dimV = n and write A = /(T). Then we have
det cT = det cA =

σ
ca
σ(1),1
ca
σ(n),n
= c
n
det A = c
n
det T
21. Let S = I and T = −I. Then we have 0 = det(S + T) ,= det S + det T = 1 + 1 = 2.
22. We can use the result of the next exercise which means that we only need to prove
this for the complex case. Let T ∈ L(C
n
) be the operator corresponding to the
matrix A. Now each block A
j
on the diagonal corresponds to some basis vectors
(e
i
, . . . , e
i+k
). These can be replaced with another set of vectors, (v
i
, . . . , v
i+k
), span-
ning the same subspace such that A
j
in this new basis is upper-triangular. We have
T(span(v
i
, . . . , v
i+k
)) ⊂ span(v
i
, . . . , v
i+k
), so after doing this for all blocks we can
assume that A is upper-triangular. The claim follows immediately.
23. This follows trivially from the formula of trace and determinant for a matrix, because
the formulas only depends on the elements of the matrix.
34
24. Let dimV = n and let A = /(T), so that B = /(T) equals the conjugate transpose
of A i.e. b
ij
= a
ji
. Then we have
det T = det A =

σ
a
σ(1),1
a
σ(n),n
=

σ
a
1,σ(1)
a
n,σ(n)
=

σ
a
1,σ(1)
a
n,σ(n)
=

σ
b
σ(1),1
b
σ(n),n
= det B = det T

.
The ﬁrst equality on the second line follows easily that every term in the upper sum
is represented by a term in the lower sum and vice versa. The second claim follows
immediately from theorem 10.31.
25. Let Ω = ¦(x, y, z) ∈ R
3
[ x
2
+ y
2
+ z
2
< 1¦ and let T be the operator T(x, y, z) =
(ax, by, cz). It’s easy to see that T(Ω) is the ellipsoid. Now Ω is a circle with radius
1. Hence
[ det T[volume(Ω) = abc
4
3
π =
4
3
πabc,
which is the volume of an ellipsoid.
35